Feedforward Neural Network Controller Research Articles

Adaptive neural control for space structure vibration suppression

Despite recent advances in efficiency, current methodologies for space structure control design still engage significant human resources for engineering development and routine maintenance. The adaptive neural control (ANC) program is part of an effort to develop neural network based controllers capable of self-optimization, on-line adaptation and autonomous fault detection and control recovery. This development in addition supports the long-term space exploration objectives for which autonomous spacecraft involving self-reliant control systems are a necessity. The ANC program comprises two phases. The first, basic phase focused on the development of efficient and completely autonomous neural network feedforward control for the case of broadband disturbances. Algorithms were developed that work with no prior modeling information about the system to be controlled and adapt to changing conditions, while minimizing or eliminating the introduction of extraneous training signals. The algorithms were demonstrated experimentally on an optical structure testbed at Harris. The second phase of the program demonstrated a more complex neural controller on the advanced space structures technology research experiments (ASTREX) test facility at the Air Force Research Laboratory capable of the fault-tolerant adaptive control of multiple sensors and actuators. This system used six actuation channels of the existing ACESA struts on the ASTREX structure to simultaneously cancel three independent tonal disturbances in the 10-15 Hz band, measured at non-collocated sensors on the secondary tower of the structure. The system demonstrated impressive fault-recovery performance, maintaining good cancellation performance with successive actuators disabled. Cancellation of individual tones was between 25 and 55 dB, with over 27 dB attenuation realized root mean square. The algorithm required very low computational throughput, operating at a sample rate of 1/20 Hz. The results of the ANC program show that adaptive cancellation systems can reduce vibrations in precision structures without prior modeling information and can adapt successfully to certain failures in actuators or sensors, optimally reconfiguring themselves without human intervention. These capabilities should significantly reduce the expense of designing and maintaining vibration control systems for spacecraft.

A diagonal recurrent neural network-based hybrid direct adaptive SPSA control system

A direct adaptive simultaneous perturbation stochastic approximation (DA SPSA) control system with a diagonal recurrent neural network (DRNN) controller is proposed. The DA SPSA control system with DRNN has simpler architecture and parameter vector size that is smaller than a feedforward neural network (FNN) controller. The simulation results show that it has a faster convergence rate than FNN controller. It results in a steady-state error and is sensitive to SPSA coefficients and termination condition. For trajectory control purpose, a hybrid control system scheme with a conventional PID controller is proposed.

Additive feedforward control with neural networks

This paper demonstrates a method to control a nonlinear, multivariable, noisy process using trained neural networks. The basis for the method is a trained neural network controller acting as the inverse process model. A training method for obtaining such an inverse process model is applied. A suitable 'shaped' (low-pass filtered) reference is used to overcome problems with excessive control action when using a controller acting as the inverse process model. The control concept is Additive Feedforward Control, where the trained neural network controller, acting as the inverse process model, is placed in a supplementary pure feedforward path to an existing feedback controller. This concept benefits from the fact, that an existing, traditional designed, feedback controller can be retained without any modifications, and after training the connection of the neural network feedforward controller to the feedback controlled process may happen gradually and controlled.

Real-Time Water Treatment Process Control with Artificial Neural Networks

The coagulation, flocculation, and sedimentation processes involve many complex physical and chemical phenomena and thus are difficult to model for process control with traditional methods. Proposed is the use of a neural network process control system for the coagulation, flocculation, and sedimentation processes. Presented is a review of influential control parameters and control requirements for these processes followed by the development of a feed forward neural network control scheme. A neural network process model was built based on nearly 2,000 sets of process control data. This model formed the major component of a software controller and was found to consistently predict the optimum alum and power activated carbon doses for different control actions. With minor modifications, the approach illustrated can be used for building control models for other water treatment processes.

Intelligent Control Paradigm for Dynamic Discrete EventSystem Simulation

Simulation based control of discrete event systems has been a potential approach to support decision-making in the manufacturing scenario. In this paper, a knowledge intensive simulation modelling approach for a discrete even system is investigated. Based on the proposed simulation model, a robust control mechanism is presented that is believed to add significant value to discrete event dynamic system. The algorithm utilises neural network feedforward control plus robust proportional derivative feedback control to achieve control performance and output stability. The novel simulation approach, as well as the proposed controller, is implemented in an Extend TM environment and the effectiveness and usefulness of the proposed controller are verified, industrially, in the hard disk drive assembly process, a significant component of the Singapore manufacturing economy.

INTELLIGENT CONTROL OF AN INTERNAL SUPPLY CHAIN IN THE HARD DISK DRIVE INDUSTRY

Simulation-based control of discrete event systems has been a potential approach to support decision-making in the manufacturing scenario. In this paper, a new simulation-based robust control mechanism is presented. The proposed algorithm utilizes neural network feedforward control plus robust proportional derivative feedback control to achieve significant control performance and output stability. A new learning mechanism and topological structure for neural networks and a new robust controller is investigated. The stability of the proposed controller is proved in detail. The proposed controller is implemented in an ExtendTM environment and the effectiveness and usefulness of the proposed controller are verified in the hard disk drive assembly process, a significant component of Singapore's manufacturing economy.

New skill learning paradigm using various kinds of neurons

Modeled from human neurons, various types of artificial neurons are developed and applied to the control algorithm. In this paper, the weights and structure of feedforward neural network controller are updated using new skill learning paradigm which consists of supervisory controller, chaotic neuron filter and associative memory. The pattern of system nonlinearity along the desired path is extracted while supervisory controller guarantees stability in the sense of the boundedness of the tracking error. Next the pattern is divided into small segments and encoded to bipolar codes depending on the existence of critical points. Comparing the encoded pattern with the pre-stored neural parameters and pattern pairs through the associative memory, the most similar one is obtained. Also, the chaotic neuron filter is used to add perturbation to neural parameters when the training of feedforward neural network is not successful with the pre-stored parameters. Finally the memory is updated with new successful parameters and pattern pairs. Simulation is performed for simple two-link robot in case of the slight modification of the desired trajectory.

Artificial-neural-network-based fast valving control in a power-generation system

This paper presents an artificial-neural-network-based controller to realize fast valving in a power-generation plant. A backpropagation algorithm is used to train the feedforward neural-network controller. The hardware implementation and the test results of the controller on a physical pilot-scale power system set-up are described in detail. Compared with some conventional fast valving methods applied to the same system, test results (both in a computer simulation and on a physical pilot-scale power system set-up) show that the neural-network controller has quite satisfactory generalisation capability, feasibility and reliability, as well as accuracy.

Active Learning for Nonlinear System Identification and Control

The identification and control of nonlinear systems continues to remain a challenging issue, Neural network models and controllers have often been effective in addressing the situation, However current neural network learning methods have been found to be limited in their generalisation abilities. Recent research has shown active learning methods to be effective in increasing the modelling reliability of a neural network system. An active learning agent has the ability to query its environment in order to make a selection of its training data. One approach to the implementation of active leaning is to use 'querying-by-committee'. This results in considerably reduced data collection and at the same time does not compromise the accuracy of identification. A nonlinear plant with both clean and noisy data is successfully modelled by such a technique and a feedforward neural network controller based upon such a model is demonstrated to perform effectively.

Motion Control of a Variable-Geometry Flexible Structure using Neural-Network Inverse Model and Fuzzy-Set Theory.

This paper presents a numerical method for motion control of a variable-geometry flexible beam structure. The transient response of the structure is calculated using updated nodal coordinates and internal forces, and the stiffness matrix for the beam elements in a finite element model. Using the results of various quasi-static finite element analyses, the neural-network feedforward controller learns an inverse model of the controlled flexible structure. To minimize the tracking error and to damp the vibration, the fuzzy feedback controller is combined with feedforward controller, which uses the above inverse model to estimate the sensitivity of the controlled structure. As an example, this method is applied to an arm-like flexible, lightweight beam structure which has 21 elements including two variable links, and the motion is analyzed.

Feedforward Neural-Network Controller for Ship Steering

A feedforward multilayer neural network is proposed to steer a ship to a given destination. For training and testing of the neural controller, the Nomoto's mathematical ship model, whose parameters change with the changing forward speed of the ship, is used. The neural controller is trained to steer the ship to a particular destination in a noise free environment. The trained controller is successful in controlling and guiding the ship to its destination in a noisy environment through number of gates. The performance of the neural network controller is tested for both severe process noise and measurement noise. The neural controller's ability to deal with noisy sea environments and inaccurate position measurements is very good.

Neural network-based control for the fiber placement composite manufacturing process

At McDonnell Douglas Aerospace (MDA), an artificial neural network-based control system has been developed and implemented to control laser heating for the fiber placement composite manufacturing process. This neurocontroller learns the inverse model of the process on-line to provide performance that improves with experience and exceeds that of conventional feedback control techniques. When untrained, the control system behaves as a proportional-integral (PI) controller. However, after learning from experience, the neural network feedforward control module provides control signals that greatly improve temperature tracking performance. Faster convergence to new temperature set points and reduced temperature deviation due to changing feed rate have been demonstrated on the machine. A cerebellar model articulation controller (CMAC) network is used for inverse modeling because of its rapid learning performance. This control system is implemented in an IBM-compatible 386 PC with an A/D board interface to the machine.

Control Of A Wall Climbing Robot Named Rostam

To automate nondestructive inspections of aircraft structures, a robot has been built equipped with jointed legs and suction cup feet that allow it to maneuver over the skinned surface of an aircraft exterior. ROSTAM (Robotic System for Total Aircraft Maintenance) has two prismatic degrees of freedom, two revolute degrees of freedom, and sensors for leg (suction cup) contact with the climbing surface. ROSTAM was designed with a sequential controller that runs on a computer attached via an umbilical cord. The sequential controller manipulates each joint separately in a distinct sequence of operations which moves the robot across the surface while maintaining adequate suction attachment during the process. To enhance the motion of the robot, a feed-forward neural network controller is trained to copy the control provided by the current sequential controller. The neural network takes as inputs, the same state and sensor information and outputs the same sequence of control actions as the sequential controller. This network controller is the first step in a plan to design a neural controller that can learn a viable walking procedure without copying an existing controller.